LLM API Pricing Comparison 2026: GPT, Claude, Gemini and DeepSeek by Blended Cost Per Million Tokens
The cheapest LLM API in August 2026 is DeepSeek V4-Flash at $0.14/M input and $0.28/M output, but the number that decides your gross margin is the blended rate for your own input:output mix, not any provider's headline input price.
Key takeaways
- Headline input price ranks providers wrong. Blended rate, weighted by your actual input:output split, is the only comparable number.
- Output multipliers vary hugely: Grok 4.3 charges 2x its input rate for output, Claude Opus 5 charges 5x, GPT-5.6 Sol charges 6x.
- GPT-5.4-mini and Grok 4.3 tie at about $1.50/M on an 80/20 input-heavy workload. At 20/80 they split to $3.75/M and $2.25/M.
- Caching and batching move the bill more than switching providers does: Anthropic cache hits are 0.1x base input, DeepSeek V4-Flash cache hits are $0.0028/M.
- Claude Sonnet 5 is $2/$10 through August 31, 2026 and $3/$15 from September 1. If you priced a plan off the introductory rate, your COGS rises 50% on a date you did not choose.
What do the major LLM APIs cost per million tokens in 2026?
As of August 2026, list prices per million tokens sit roughly like this:
| Provider | Model | Input $/M | Output $/M | | --- | --- | --- | --- | | OpenAI | GPT-5.6 Sol | $5.00 | $30.00 | | OpenAI | GPT-5.4 | $2.50 | $15.00 | | OpenAI | GPT-5.4-mini | $0.75 | $4.50 | | Anthropic | Claude Sonnet 5 (intro, to Aug 31) | $2.00 | $10.00 | | Anthropic | Claude Opus 5 | $5.00 | $25.00 | | Google | Gemini 3.1 Pro (under 200k) | $2.00 | $12.00 | | Google | Gemini 3.5 Flash-Lite | $0.30 | $2.50 | | DeepSeek | V4-Flash | $0.14 | $0.28 | | xAI | Grok 4.3 (under 200k) | $1.25 | $2.50 | | Mistral | Small 4 | $0.15 | $0.60 |
Rates compiled from provider pricing pages as summarised by Spheron, August 6, 2026. Always re-check the provider docs before you commit a number to a financial model.
How do you calculate a blended rate?
One line of arithmetic:
Run the same eight models across three workload shapes and the ranking moves:
| Model | 80/20 input-heavy | 50/50 | 20/80 output-heavy | | --- | --- | --- | --- | | DeepSeek V4-Flash | $0.17/M | $0.21/M | $0.25/M | | Gemini 3.5 Flash-Lite | $0.74/M | $1.40/M | $2.06/M | | Grok 4.3 | $1.50/M | $1.88/M | $2.25/M | | GPT-5.4-mini | $1.50/M | $2.63/M | $3.75/M | | Claude Sonnet 5 (intro) | $3.60/M | $6.00/M | $8.40/M | | GPT-5.4 | $5.00/M | $8.75/M | $12.50/M | | Claude Opus 5 | $9.00/M | $15.00/M | $21.00/M | | GPT-5.6 Sol | $10.00/M | $17.50/M | $25.00/M |
Grok 4.3 and GPT-5.4-mini are identical at 80/20 and 67% apart at 20/80. Nothing changed except the shape of the traffic.
Why does the output multiplier matter more than the price cut?
Here is the part most comparison posts leave out: the input:output ratio is not a fixed property of your product. It drifts, and it drifts in one direction.
A support copilot that answers in two sentences is input-heavy. Add agentic tool calls, add a drafting feature, add extended reasoning, and output share climbs quietly with every release you ship. You do not get a pricing email when that happens. You just watch gross margin sag over two quarters and blame it on growth.
Make it concrete. Say a user consumes 2M tokens a month and you charge $29 a seat:
- Input-heavy on GPT-5.4 at $5.00/M blended: about $10.00 of COGS per user, roughly 66% gross margin.
- Output-heavy on the same model at $12.50/M blended: about $25.00 of COGS per user, roughly 14% gross margin.
Same provider, same price, same customer. The only thing that moved was the ratio. Figures are illustrative, but the shape of the result is not.
Which provider is cheapest for which workload?
- Chat and short turns: the rows that stay cheap without discounts. DeepSeek V4-Flash, Gemini 3.5 Flash-Lite, Mistral Small 4, GPT-5.4-mini.
- Long context over 200k tokens: Gemini 3.1 Pro and both Grok tiers roughly double past 200k. Claude's Opus and Sonnet 5 include the full 1M context at standard rate, so a model that looks expensive on a short prompt can win on a long one.
- Batch and async: OpenAI, Anthropic and Google all offer 50% off batch. Stack Anthropic caching on top and the input portion can fall by up to 95% against uncached synchronous pricing. Output tokens only get the flat batch discount, so blended savings land lower than the input-only headline.
What should you actually do this week?
Pull the real input and output token counts out of production logs for your top three endpoints. Compute the blended rate per endpoint, not per product. Then re-run it at the ratio you expect after your next two features ship, because that is the number your pricing has to survive. You can model both scenarios, and the September 1 Sonnet 5 step, in Calcaas in a few minutes.
The cheapest API is not a fact about the market. It is a fact about your traffic.
Frequently asked questions
What is the cheapest LLM API in 2026?
By list price, DeepSeek V4-Flash at $0.14/M input and $0.28/M output, which blends to about $0.17/M on an 80/20 input-heavy workload. Cache hits drop input to $0.0028/M. DeepSeek has announced a peak/off-peak policy that would double prices during Beijing business hours, so check current status before budgeting.
How do you calculate cost per million tokens for an LLM API?
Blended cost = (input share x input price) + (output share x output price). For an 80/20 workload on GPT-5.4 at $2.50/M input and $15/M output: (0.8 x 2.50) + (0.2 x 15.00) = $5.00/M. Use your own ratio from production logs, not a generic assumption.
Why do GPT, Claude and Gemini rank differently depending on the task?
Because each prices output at a different multiple of input. Grok 4.3 charges 2x, Claude Opus 5 charges 5x, GPT-5.6 Sol charges 6x. The flatter the multiplier, the better the model looks as output share rises, regardless of headline input price.
Does switching providers fix a thin AI gross margin?
Usually not on its own. Caching, batching and trimming retrieved context often move the bill further than a provider swap, and they do not cost you an evaluation cycle. Switch providers when the blended rate gap survives those optimisations. Place the JSON-LD above inside a <script type="application/ld+json"> tag in the page head. Source article: https://www.spheron.network/blog/llm-api-pricing-comparison-gpt-claude-gemini-deepseek-2026/
More from the blog
LLM EconomicsOpen-Weight vs Frontier LLMs: How to Model the Real Cost Difference
Open-weight models are close enough to frontier quality that the switch is now a pricing decision, but the per-token sticker price hides the serving, safety and compliance work the API price was quietly covering.
LLM EconomicsInference Engineering Is a Gross Margin Lever, Not an Infra Detail
Inference engineering is the discipline of turning model weights into a fast, affordable production API, and the throughput gains it produces land in someone's gross margin: yours if you serve the model, your provider's if you buy tokens.
Open-Weight Models and Your LLM Cost Per Token: What DeepSeek V4 Flash Actually Changes
An open-weight model landing within a few points of the frontier does not automatically cut your AI bill: your real cost per token is set by how many tokens you consume, how fast you can serve them, and whether you own the hardware at all.
Pricing math, in your inbox.
One short note a week on AI pricing, token economics, and margin. No spam, unsubscribe anytime.